Slides for https://youtu.be/hGNycrko5kc

1. Introduction to Transformers
for NLP
where we are and how we got here
Olga Petrova
AI Product Manager
DataTalks.Club

2. Preliminaries
● Who I am:
○ Product Manager for AI PaaS at Scaleway
Scaleway: European cloud provider, originating from
France

3. Preliminaries
● Who I am:
○ Product Manager for AI PaaS at Scaleway
Scaleway: European cloud provider, originating from
France
○ Machine Learning engineer at Scaleway

4. Preliminaries
● Who I am:
○ Product Manager for AI PaaS at Scaleway
Scaleway: European cloud provider, originating from
France
○ Machine Learning engineer at Scaleway
○ Quantum physicist at École Normale Supérieure, the
Max Planck Institute, Johns Hopkins University

5. Preliminaries
● Who I am:
○ Product Manager for AI PaaS at Scaleway
Scaleway: European cloud provider, originating from
France
○ Machine Learning engineer at Scaleway
○ Quantum physicist at École Normale Supérieure, the
Max Planck Institute, Johns Hopkins University
www.olgapaints.net

6. Outline
1. NLP 101: how do we represent words in machine learning
🤖 Tokenization 🤖 Word embeddings
2. Order matters: Recurrent Neural Networks
🤖 RNNs 🤖 Seq2Seq models
🤖 What is wrong with RNNs
3. Attention is all you need!
🤖 What is Attention 🤖 Attention, the linear algebra perspective
4. The Transformer architecture
🤖 Mostly Attention
5. BERT: contextualized word embeddings

7. Outline
1. NLP 101: how do we represent words in machine learning
🤖 Tokenization 🤖 Word embeddings
2. Order matters: Recurrent Neural Networks
🤖 RNNs 🤖 Seq2Seq models
🤖 What is wrong with RNNs
3. Attention is all you need!
🤖 What is Attention 🤖 Attention, the linear algebra perspective
4. The Transformer architecture
🤖 Mostly Attention
5. BERT: contextualized word embeddings

8. Representing words
NLP 101 RNNs Attention Transformers BERT

9. Representing words
I play with my dog.
NLP 101 RNNs Attention Transformers BERT

10. Representing words
I play with my dog.
● Machine learning: inputs are vectors / matrices
NLP 101 RNNs Attention Transformers BERT

11. Representing words
I play with my dog.
● Machine learning: inputs are vectors / matrices
ID Age Gender Weight Diagnosis
264264 27 0 72 1
908696 61 1 80 0
NLP 101 RNNs Attention Transformers BERT

12. Representing words
ID Age Gender Weight Diagnosis
264264 27 0 72 1
908696 61 1 80 0
I play with my dog.
● Machine learning: inputs are vectors / matrices
● How do we convert “I play with my dog” into numerical values?
NLP 101 RNNs Attention Transformers BERT

13. Representing words
I play with my dog.
I
play
with
my
dog
.
NLP 101 RNNs Attention Transformers BERT

14. Representing words
I play with my dog.
I
play
with
my
dog
.
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
1
0
0
0
0
1
0
0
● One hot encoding
● Dimensionality = vocabulary size
1
0
0
0
0
0
1
2
3
4
5
6
NLP 101 RNNs Attention Transformers BERT

15. Representing words
I play with my dog.
I play with my cat.
I
play
with
my
dog
.
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
1
0
0
0
0
1
0
0
● One hot encoding
● Dimensionality = vocabulary size
1
2
3
4
5
6
NLP 101 RNNs Attention Transformers BERT

16. Representing words
I play with my dog.
I play with my cat.
I
play
with
my
dog
.
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
1
0
0
0
0
0
1
0
0
0
● One hot encoding
● Dimensionality = vocabulary size
cat
0
0
0
0
0
0
1
1
2
3
4
5
6
7
NLP 101 RNNs Attention Transformers BERT

17. play played playing
I play / played / was playing with my dog.
Word = token: , ,
Subword = token: , ,
E.g.: played = ;
● In practice, subword tokenization is used most often
● For simplicity, we’ll go with word = token for the rest of the talk
Tokenization
play #ed #ing
play #ed
NLP 101 RNNs Attention Transformers BERT

18. Word embeddings
= …… N = …… = …… = ……
One hot encodings N = vocabulary size
● puppy : dog = kitten : cat. How can we record this?
● Goal: a lower-dimensional representation of a word (M << N)
● Words are forced to “group together” in a meaningful way
0
1
cat
0
0
dog
0
0
1
0
0
0
0
1
kitten
0
0
1
0
puppy
NLP 101 RNNs Attention Transformers BERT

19. Word embeddings
= …… N = …… = …… = ……
One hot encodings
● dog : puppy = cat : kitten. How can we record this?
● Goal: a lower-dimensional representation of a word (M << N)
● Words are forced to “group together” in a meaningful way
0
1
cat
0
0
dog
0
0
1
0
0
0
0
1
kitten
0
0
1
0
puppy
NLP 101 RNNs Attention Transformers BERT

20. Word embeddings
= …… N = …… = …… = ……
One hot encodings
● dog : puppy = cat : kitten. How can we record this?
● Goal: a lower-dimensional representation of a word (M << N)
● Words are forced to “group together” in a meaningful way
0
1
cat
0
0
dog
0
0
1
0
0
0
0
1
kitten
0
0
1
0
puppy
NLP 101 RNNs Attention Transformers BERT

21. Word embeddings
= …… N = …… = …… = ……
One hot encodings
● dog : puppy = cat : kitten. How can we record this?
● Goal: a lower-dimensional representation of a word (M << N)
● Words are forced to “group together” in a meaningful way
0
1
cat
0
0
dog
0
0
1
0
0
0
0
1
kitten
0
0
1
0
puppy
1
0
1
0
1
1
1
0
0
0
1
0
= M = = =
cat dog kitten puppy
NLP 101 RNNs Attention Transformers BERT

22. Word embeddings
What do you mean, “meaningful way”?
= …… N = …… = …… = ……
One hot encodings
● dog : puppy = cat : kitten. How can we record this?
● Goal: a lower-dimensional representation of a word (M << N)
● Words are forced to “group together” in a meaningful way
1
0
1
0
1
1
1
0
0
0
1
0
= M = = =
cat dog kitten puppy
NLP 101 RNNs Attention Transformers BERT

23. Word embeddings
What do you mean, “meaningful way”?
= …… N = …… = …… = ……
One hot encodings
● dog : puppy = cat : kitten. How can we record this?
● Goal: a lower-dimensional representation of a word (M << N)
● Words are forced to “group together” in a meaningful way
1
0
1
0
1
1
1
0
0
0
1
0
= M = = =
cat dog kitten puppy
1 for feline
1 for canine - = -
1 for adult / 0 for baby
dog puppy cat kitten
NLP 101 RNNs Attention Transformers BERT

24. How do we arrive at this representation?
Word embeddings
NLP 101 RNNs Attention Transformers BERT

25. How do we arrive at this representation?
● Pass the one-hot vector through a network with a bottleneck layer
Word embeddings
NLP 101 RNNs Attention Transformers BERT

26. How do we arrive at this representation?
● Pass the one-hot vector through a network with a bottleneck layer
= multiply the N-dim vector by NxM matrix to get an M-dim vector
Word embeddings
Embedding matrix W
One-hot vector
Embedded
vector
NLP 101 RNNs Attention Transformers BERT

27. How do we arrive at this representation?
● Pass the one-hot vector through a network with a bottleneck layer
= multiply the N-dim vector by NxM matrix to get an M-dim vector
● Some information gets lost in the process
Word embeddings
Embedding matrix W
One-hot vector
Embedded
vector
NLP 101 RNNs Attention Transformers BERT

28. How do we arrive at this representation?
● Pass the one-hot vector through a network with a bottleneck layer
= multiply the N-dim vector by NxM matrix to get an M-dim vector
● Some information gets lost in the process
● Embedding matrix: through machine learning!
Word embeddings
Embedding matrix W
One-hot vector
Embedded
vector
In practice:
● Embedding can be part of your
model that you train for a ML task
● Or: use pre-trained embeddings
(e.g. Word2Vec, GloVe, etc)
NLP 101 RNNs Attention Transformers BERT

29. Tokenization:
blog.floydhub.com/tokenization-nlp/
Word embeddings:
jalammar.github.io/illustrated-word2vec/
Further reading
NLP 101 RNNs Attention Transformers BERT

30. Outline
1. NLP 101: how do we represent words in machine learning
🤖 Tokenization 🤖 Word embeddings
2. Order matters: Recurrent Neural Networks
🤖 RNNs 🤖 Seq2Seq models
🤖 What is wrong with RNNs
3. Attention is all you need!
🤖 What is Attention 🤖 Attention, the linear algebra perspective
4. The Transformer architecture
🤖 Mostly Attention
5. BERT: contextualized word embeddings

31. Sequence to Sequence models
● Not all seq2seq models are NLP
● Not all NLP models are seq2seq!
● But: they are common enough
NLP 101 RNNs Attention Transformers BERT

32. Sequence to Sequence models
● Not all seq2seq models are NLP
● Not all NLP models are seq2seq!
● But: they are common enough
NLP 101 RNNs Attention Transformers BERT
ML model Sequence 2
Sequence 1

33. Sequence to Sequence models
● Not all seq2seq models are NLP
● Not all NLP models are seq2seq!
● But: they are common enough
NLP 101 RNNs Attention Transformers BERT
ML model Sequence 2
Sequence 1
● Language modeling, question answering, machine translation, etc
Machine
Translation
I am a student
je suis étudiant

34. Sequence to Sequence models
● Not all seq2seq models are NLP
● Not all NLP models are seq2seq!
● But: they are common enough
NLP 101 RNNs Attention Transformers BERT
ML model Sequence 2
Sequence 1
● Language modeling, question answering, machine translation, etc
Machine
Translation
I am a student
je suis étudiant
Takeaways:
1. Multiple elements in a sequence
2. Order of the elements matters

35. RNNs: Recurrent Neural Networks
NLP 101 RNNs Attention Transformers BERT

36. RNNs: Recurrent Neural Networks
What does this mean???
NLP 101 RNNs Attention Transformers BERT

37. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT

38. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT

39. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT

40. RNNs: the seq2seq example
Output sequence
Input sequence
NLP 101 RNNs Attention Transformers BERT

41. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Inputs (word embeddings)

42. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Inputs (word embeddings)

43. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Inputs (word embeddings)

44. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Inputs (word embeddings)

45. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Inputs (word embeddings)

46. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Inputs (word embeddings)

47. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Context vector: contains the
meaning of the whole input
sequence
Inputs (word embeddings)

48. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Context vector: contains the
meaning of the whole input
sequence
Inputs (word embeddings)
Temporal direction (remember the Wikipedia
definition?)

49. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Context vector: contains the
meaning of the whole input
sequence
Inputs (word embeddings)
Temporal direction (remember the Wikipedia
definition?)

50. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Inputs (word embeddings)

51. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Inputs (word embeddings)

52. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Inputs (word embeddings)

53. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Inputs (word embeddings)

54. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Inputs (word embeddings)

55. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Inputs (word embeddings)

56. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Inputs (word embeddings)

57. RNNs: the seq2seq example
NLP 101 RNNs Attention Transformers BERT
Outputs (one-hot)
Inputs (word embeddings)

58. Does this really work?
NLP 101 RNNs Attention Transformers BERT

59. Does this really work?
Yes! (For the most part) Google Translate has been using these since ~ 2016
NLP 101 RNNs Attention Transformers BERT

60. Does this really work?
Yes! (For the most part) Google Translate has been using these since ~ 2016
What is the catch?
NLP 101 RNNs Attention Transformers BERT

61. Does this really work?
Yes! (For the most part) Google Translate has been using these since ~ 2016
What is the catch?
● Catch # 1: Recurrent = Non-parallelizable 🐌
NLP 101 RNNs Attention Transformers BERT

62. Does this really work?
Yes! (For the most part) Google Translate has been using these since ~ 2016
What is the catch?
● Catch # 1: Recurrent = Non-parallelizable 🐌
● Catch # 2: Relying on a fixed-length context vector to capture the full input
sequence. What can go wrong?
NLP 101 RNNs Attention Transformers BERT

63. Does this really work?
Yes! (For the most part) Google Translate has been using these since ~ 2016
What is the catch?
● Catch # 1: Recurrent = Non-parallelizable 🐌
● Catch # 2: Relying on a fixed-length context vector to capture the full input
sequence. What can go wrong?
“I am a student.”
NLP 101 RNNs Attention Transformers BERT

64. Does this really work?
Yes! (For the most part) Google Translate has been using these since ~ 2016
What is the catch?
● Catch # 1: Recurrent = Non-parallelizable 🐌
● Catch # 2: Relying on a fixed-length context vector to capture the full input
sequence. What can go wrong?
“I am a student.” ✅
NLP 101 RNNs Attention Transformers BERT

65. Does this really work?
Yes! (For the most part) Google Translate has been using these since ~ 2016
What is the catch?
● Catch # 1: Recurrent = Non-parallelizable 🐌
● Catch # 2: Relying on a fixed-length context vector to capture the full input
sequence. What can go wrong?
“I am a student.” ✅
“It was a wrong number that started it, the telephone ringing three times in the dead
of night, and the voice on the other end asking for someone he was not.”
NLP 101 RNNs Attention Transformers BERT

66. Does this really work?
Yes! (For the most part) Google Translate has been using these since ~ 2016
What is the catch?
● Catch # 1: Recurrent = Non-parallelizable 🐌
● Catch # 2: Relying on a fixed-length context vector to capture the full input
sequence. What can go wrong?
“I am a student.” ✅
“It was a wrong number that started it, the telephone ringing three times in the dead
of night, and the voice on the other end asking for someone he was not.” 😱
NLP 101 RNNs Attention Transformers BERT

67. Outline
1. NLP 101: how do we represent words in machine learning
🤖 Tokenization 🤖 Word embeddings
2. Order matters: Recurrent Neural Networks
🤖 RNNs 🤖 Seq2Seq models
🤖 What is wrong with RNNs
3. Attention is all you need!
🤖 What is Attention 🤖 Attention, the linear algebra perspective
4. The Transformer architecture
🤖 Mostly Attention
5. BERT: contextualized word embeddings

68. The Attention Mechanism
NLP 101 RNNs Attention Transformers BERT

69. The Attention Mechanism
NLP 101 RNNs Attention Transformers BERT
Recurrent NN from the previous section

70. The Attention Mechanism
NLP 101 RNNs Attention Transformers BERT
Recurrent NN from the previous section

71. The Attention Mechanism
NLP 101 RNNs Attention Transformers BERT
Recurrent NN from the previous section
Send all of the hidden states to the Decoder, not just
the last one

72. The Attention Mechanism
NLP 101 RNNs Attention Transformers BERT
Recurrent NN from the previous section
Send all of the hidden states to the Decoder, not just
the last one
The Decoder determines which input elements get
attended to via a softmax layer

73. Further reading
NLP 101 RNNs Attention Transformers BERT
jalammar.github.io/illustrated-transformer/
My blog post that this talk is based on + the linear algebra perspective on
Attention:
towardsdatascience.com/natural-language-processing-the-age-of-transformers-a3
6c0265937d

74. Outline
1. NLP 101: how do we represent words in machine learning
🤖 Tokenization 🤖 Word embeddings
2. Order matters: Recurrent Neural Networks
🤖 RNNs 🤖 Seq2Seq models
🤖 What is wrong with RNNs
3. Attention is all you need!
🤖 What is Attention 🤖 Attention, the linear algebra perspective
4. The Transformer architecture
🤖 Mostly Attention
5. BERT: contextualized word embeddings

75. Attention is All You Need © Google, 2017
NLP 101 RNNs Attention Transformers BERT

76. Attention is All You Need © Google, 2017
● Keep the Attention mechanism, throw away the sequential RNN
● Reminder: the basic idea of Attention was to pass all of the hidden states as inputs
NLP 101 RNNs Attention Transformers BERT

77. Attention is All You Need © Google, 2017
● Keep the Attention mechanism, throw away the sequential RNN
● Reminder: the basic idea of Attention was to pass all of the hidden states as inputs
● Transformer:
○ Decoder and Encoder both have multiple layers
NLP 101 RNNs Attention Transformers BERT

78. Attention is All You Need © Google, 2017
● Keep the Attention mechanism, throw away the sequential RNN
● Reminder: the basic idea of Attention was to pass all of the hidden states as inputs
● Transformer:
○ Decoder and Encoder both have multiple layers
○ The output hidden states of the Encoder are inputs to all the layers of the Decoder
NLP 101 RNNs Attention Transformers BERT

79. Attention is All You Need © Google, 2017
● Keep the Attention mechanism, throw away the sequential RNN
● Reminder: the basic idea of Attention was to pass all of the hidden states as inputs
● Transformer:
○ Decoder and Encoder both have multiple layers
○ The output hidden states of the Encoder are inputs to all the layers of the Decoder
○ The intermediate hidden states of the Encoder are also inputs… to other sequence elements’
Encoder layers! Self-Attention
NLP 101 RNNs Attention Transformers BERT

80. Attention is All You Need © Google, 2017
● Keep the Attention mechanism, throw away the sequential RNN
● Reminder: the basic idea of Attention was to pass all of the hidden states as inputs
● Transformer:
○ Decoder and Encoder both have multiple layers
○ The output hidden states of the Encoder are inputs to all the layers of the Decoder
○ The intermediate hidden states of the Encoder are also inputs… to other sequence elements’
Encoder layers! Self-Attention
○ Also Self-Attention in the Decoder
NLP 101 RNNs Attention Transformers BERT

81. Attention all around
NLP 101 RNNs Attention Transformers BERT

82. Attention all around
NLP 101 RNNs Attention Transformers BERT
Word embedding + position encoding

83. Attention all around
● All of the sequence elements (w1, w2,
w3) are being processed in parallel
● Each element’s encoding gets
information (through hidden states)
from other elements
● Bidirectional by design 😍
NLP 101 RNNs Attention Transformers BERT

84. Why Self-Attention?
● “The animal did not cross the road
because it was too tired.”
vs.
● “The animal did not cross the road
because it was too wide.”
NLP 101 RNNs Attention Transformers BERT

85. Why Self-Attention?
● “The animal did not cross the road
because it was too tired.”
vs.
● “The animal did not cross the road
because it was too wide.”
Uni-directional RNN would have missed
this! Bi-directional is nice.
NLP 101 RNNs Attention Transformers BERT

87. Outline
1. NLP 101: how do we represent words in machine learning
🤖 Tokenization 🤖 Word embeddings
2. Order matters: Recurrent Neural Networks
🤖 RNNs 🤖 Seq2Seq models
🤖 What is wrong with RNNs
3. Attention is all you need!
🤖 What is Attention 🤖 Attention, the linear algebra perspective
4. The Transformer architecture
🤖 Mostly Attention
5. BERT: contextualized word embeddings

88. (Contextual) word embeddings
● Dense (i.e. not sparse) vectors capturing the meaning of tokens/words
NLP 101 RNNs Attention Transformers BERT

89. (Contextual) word embeddings
● Dense (i.e. not sparse) vectors capturing the meaning of tokens/words
● Meaning often depends on the context (surrounding words)
NLP 101 RNNs Attention Transformers BERT

90. (Contextual) word embeddings
● Dense (i.e. not sparse) vectors capturing the meaning of tokens/words
● Meaning often depends on the context (surrounding words)
● Self-Attention in Transformers processes context when encoding a token
NLP 101 RNNs Attention Transformers BERT

91. (Contextual) word embeddings
● Dense (i.e. not sparse) vectors capturing the meaning of tokens/words
● Meaning often depends on the context (surrounding words)
● Self-Attention in Transformers processes context when encoding a token
● Can we just take the Transformer Encoder and use it to embed words?
NLP 101 RNNs Attention Transformers BERT

92. (Contextual) word embeddings
● Dense (i.e. not sparse) vectors capturing the meaning of tokens/words
● Meaning often depends on the context (surrounding words)
● Self-Attention in Transformers processes context when encoding a token
● Can we just take the Transformer Encoder and use it to embed words?
NLP 101 RNNs Attention Transformers BERT

93. (Contextual) word embeddings
BERT: Bidirectional Encoder Representations from Transformers
NLP 101 RNNs Attention Transformers BERT

94. (Contextual) word embeddings
BERT: Bidirectional Encoder Representations from Transformers
Instead of Word2Vec, use pre-trained BERT for your NLP tasks ;)
NLP 101 RNNs Attention Transformers BERT

95. What was BERT pre-trained on?
Task # 1: Masked Language Modeling
● Take a text corpus, randomly mask 15% of words
● Put a softmax layer (dim = vocab size) on top of BERT encodings for the words
that are present to predict which word is missing
NLP 101 RNNs Attention Transformers BERT

96. What was BERT pre-trained on?
Task # 1: Masked Language Modeling
● Take a text corpus, randomly mask 15% of words
● Put a softmax layer (dim = vocab size) on top of BERT encodings for the words
that are present to predict which word is missing
Task # 2: Next Sentence Prediction
● Input sequence: two sentences separated by a special token
● Binary classification: are these sentences successive or not?
NLP 101 RNNs Attention Transformers BERT

97. Further reading and coding
My blog posts:
towardsdatascience.com/natural-language-processing-the-age-of-transformers-a36c0
265937d
blog.scaleway.com/understanding-text-with-bert/
The HuggingFace 🤗 Transformers library: huggingface.co/
NLP 101 RNNs Attention Transformers BERT